PR for AI Search: How Agencies Earn Citations in ChatGPT Answers

A PR agency gets a client cited in AI answers by earning coverage on the sources those answers already cite for the client's topics, which you find by testing real prompts and recording the cited pages. Nobody can guarantee a citation, so pitch for the sources the test shows, then re-test.

That order, test first and pitch second, is the whole difference between guessing and doing PR for AI search. The strategy articles on this topic mostly say "get coverage in authoritative outlets", which is true and not actionable. Which outlets, for which questions, in whose answers? The method below answers that for one client in about a week, and the tracking sheet, the pitching tactics and the reporting wording all follow from its results.

Follow me on Instagram@sagnikteaches

What is verified and what is only observed

Keep two lists, and never mix them in a client report.

Connect on LinkedInSagnik Bhattacharya

Documented by the vendors. OpenAI's help page says ChatGPT can search the web and that responses using search may include citations, with a Sources view for the cited pages. It also says results and citations can be incomplete, outdated or incorrect. To make a website eligible for inclusion in ChatGPT search, OpenAI says the site must allow its search crawler, OAI-SearchBot, and the host or content delivery network must allow traffic from OpenAI's published search-bot IP addresses; it adds that placement is not guaranteed. Google's Search Central guide to succeeding in its generative AI features (AI Overviews and AI Mode) says its generative AI features are rooted in its core Search ranking and quality systems and that good SEO practice remains relevant. It also warns that producing separate content for every query variation primarily to manipulate AI responses can violate its spam policy on scaled content abuse.

Subscribe on YouTube@sagnikteaches

Only observed. Which publications get cited most is not published by the vendors. It varies by topic, product, location and the day you ask. Anything you learn about "what AI likes" comes from your own tests, and you should present it as such: "In our test of 60 answers for this client's category, ...".

That distinction protects you. When a client asks why a competitor is cited and they are not, you can show a method and a result, rather than a theory.

The citation test, step by step

Step 1: Write the prompt list (45 minutes)

Write 15 to 25 prompts a real customer might type, in three groups:

• Category questions: "what's the best way to ...", "how do I choose a ...".
• Comparison and recommendation prompts: "best [category] for [situation]", "alternatives to [competitor]".
• Brand prompts: "is [client] any good", "[client] vs [competitor]".

Write them in natural language, the way a person talks to an assistant. Ask the client's sales or support team for the last ten questions customers actually asked. Those beat anything you invent.

Step 2: Run each prompt three times (2-3 hours)

Use a logged-out or clean browser session where possible, with search switched on, and run each prompt on at least two assistants. The three runs matter because answers change from run to run. Paste each answer and its list of cited sources into a spreadsheet: one row per citation.

Step 3: Classify the sources (45 minutes)

Give each cited domain a type: owned (the client's site), national press, trade or specialist press, local news, review or directory site, forum or community, non-profit or public body, competitor, other. You can ask an AI tool to propose the type from a list of domains, then check any you do not recognise. That is a good use for it: the classification is simple and a wrong label is easy to spot. Do the counting in a spreadsheet, not in the chat.

Step 4: Count and read the pattern (30 minutes)

Count three things per prompt group: how often the client is named, how often each competitor is named, and the distribution of source types. Then read, rather than count, the five most-cited pages. Why does an answer lean on that page? Often because it is a clear comparison, a recent list, original data or a plain explanation. That tells you what kind of coverage to pitch.

Worked example: a garden centre group, 20 prompts, three runs

Say the client is a garden centre group with a schools scheme. You run 20 prompts three times on ChatGPT with search on, producing 60 answers and, say, 310 citations; the second assistant's rows go in their own tab and get the same treatment. All figures here are illustrative, to show how the method reads.

What the pattern suggests, and what it does not. Local news and specialist press together supply almost half the citations, but the client is rarely named in those answers. Non-profit bodies are cited for school-garden questions and never mention the client. That points to two actions: pitch local news and gardening trade titles with something citable (a figure, a how-to or a local story), and pursue a mention or partnership listing with the education charity whose guide is cited. Whether the actions work is a separate question, which the re-test answers.

What it does not show: that those sources "cause" citations, or that the next test will look the same. Present it as "where the answers look for this topic today".

Coverage worth pitching once the test is in

The tactics below follow from the cited-source types above. They are ordinary digital PR, pointed at sources the test shows. If you build links by outreach, the link outreach tutorial has the email side.

1. A citable fact from the client. First-party numbers (a sales trend, a survey the client ran, a count of something they know better than anyone) give journalists a reason to quote them and give answer engines a specific statement to attribute. Publish the full figures on the client's own site, so the coverage can link to a stable page.
2. Expert commentary tied to a named person. A quoted spokesperson with a clear job title and a bio page makes it easy for a source to describe who is speaking. Keep the bio page factual.
3. Inclusion in the lists and comparisons already cited. If the test shows a trade title's "best of" list gets cited, pitch the writer for inclusion on their next update, with a reason that is true.
4. Review and directory profiles. If review sites are cited, make sure the client's profile is accurate, complete and has recent genuine reviews. Never buy reviews.
5. Clear explanatory content on the client's own site. Google's guide points to unique, non-commodity content and clear structure. A page that answers a customer question better than anyone else is more likely to be quoted.

Pair this with the basics: the client's pages must be crawlable and indexed. A blocked search crawler is the most common avoidable reason a good page never appears, and the fix is for the client's web team to review the robots settings. On the Google side there is one more switch to check. Search Console has a Search generative AI control; the default lets a site's content appear in AI Overviews and AI Mode, and Google's guide says a site must be included there to be eligible. If someone on the client side chose to exclude the site, coverage pointing at its pages cannot surface them in those features, so ask the web team to confirm the setting before you start the test.

A one-page tracking sheet for each client

Client: [name]            Test date: [date]            Assistant(s): [names]
Prompt set: 20 prompts x 3 runs = 60 answers

Headline numbers (from the sheet, not from the chat):
- Answers naming the client:           [n] of 60
- Answers naming the top competitor:   [n] of 60
- Citations to owned pages:            [n] of [total]
- Top five cited domains:              1. ... 2. ... 3. ... 4. ... 5. ...

Source types most cited:               [type] [share], [type] [share]
Gaps (prompts where the client never appears): [list]
Coverage achieved since last test:     [list of pieces and dates]
Next test date: [date, four to six weeks on]

Re-test every four to six weeks with the same prompts, so changes are comparable, and add a few new ones each time. Answers change when the models and search indexes change, so a drop between two tests does not always mean something you did. Record the date, the assistants and any known product changes, to keep the log honest. For how the same idea applies to a company's own site, the tutorial on tracking AI search customers covers referral traffic and surveys.

Using AI to run the test

An assistant can speed the clerical parts of this. Give it the pasted answers and sources, and ask for the table:

Below are pasted answers from an AI assistant, each followed by its cited sources.
For every cited source, produce one row:
Prompt | Domain | Source type (from this list: Owned, National press,
Trade press, Local news, Review/directory, Forum, Non-profit/public body,
Competitor, Other) | Is the client named in the answer? (Yes/No)
Client: [name]. Competitors: [list].
Do not summarise. Do not add sources. If a domain is unfamiliar, put "Other".

A realistic fault to expect: the tool sometimes marks a regional news site as "National press" because of its name. You catch this by sorting the table by domain and scanning the source types once. Another: it may not capture every citation if the pasted text omitted a Sources list, so paste the sources view each time and not only the answer text.

When the answer gets the client wrong

Your test will also reveal something clients care about more than citation counts: what the assistants say about them, and whether it is right. Read every answer that names the client and note errors in a separate column: wrong opening hours, a discontinued product, an old price, a competitor's scheme credited to the client.

Most of these errors trace back to a cited page. The fix is rarely to argue with the assistant. It is to correct the source: update the client's own page, ask a directory to fix a listing, or ask the publication to amend a stale article. That is traditional PR work, and it is the most tangible win you can show in the first month. Illustratively, a test might find that three of 60 answers describe the schools scheme as "paid", citing a two-year-old local news piece. A polite correction request to the publication, with a link to the current scheme page, fixes the source and gives you a before-and-after to show.

Always check a cited source before you describe it to the client. OpenAI's own help page advises opening a cited source to check that it supports the answer, which is a good habit for your whole team.

Choosing which topics to test first

A client with fifty services cannot be tested on all of them. Choose by value and by fit:

1. The three services or products that bring in most revenue.
2. Questions customers ask before buying, since those are where a recommendation matters.
3. Topics where the client has real expertise, a story or data nobody else has. Those are the ones you can win with coverage rather than volume.

Leave out anything the client cannot actually serve. A prompt about a service they do not offer produces citations that tell you nothing useful.

What Google says about mentions you create for the purpose

Google's guide to generative AI features includes a short recap telling site owners they can ignore tactics such as chunking content, creating unnecessary AI text files, or pursuing inauthentic mentions. For a PR agency, the last one is the line to repeat to clients who ask for fast results. Paid placements disguised as editorial, fake reviews, and mentions seeded in forums by staff are the sort of tactic that may produce a short-term bump and a long-term problem. Earned coverage that a journalist chose to publish is slower, and it is also the work your agency exists to do.

A first-week plan

That is about seven hours for a first test, which you can price as a fixed-fee starter project. An illustrative sum: at an internal cost of $60 an hour, seven hours is $420 of time. If a typical retainer is several times that, the test is a sensible first deliverable. Later tests, with the prompts and sheet in place, should take around three hours. These numbers depend on your team, so time your first run and revise.

A client with very little coverage to build on

A new or small client may be absent from every answer, and the test will show only competitors and neutral sources. That is still useful, as it gives a map of where to start. Begin with the owned and easiest sources: an accurate, complete site; consistent business listings; accurate profiles on review and directory sites. Then add one citable piece of original information a journalist could use, and pitch two or three outlets from the test's top list. Set the expectation that results come over months, and that you will re-test every four to six weeks. Resist the urge to promise a result in the first test cycle; the honest version is "we will know more after the second test".

Why one assistant is not enough

Different assistants search differently and cite different pages, so a result from one cannot be assumed for another. Test at least two, such as ChatGPT with search on and Google's AI Mode, and keep their rows separate in the sheet. Where they agree on a source, that source is a strong target. Where they diverge, report both and say so. Google's Search Console also has a Generative AI performance report, which shows the client's own impressions in AI Overviews and AI Mode over time and by page, device and country. That is a second, independent view of visibility on the Google side; it counts impressions of the client's pages rather than the third-party pieces that answers cite, so use it as a cross-check, not a replacement for the test.

Claims to avoid in pitches and client reports

• "We guarantee citations in ChatGPT." Nobody can, for the reasons above.
• "This coverage caused the citation." You can say the coverage appears among cited sources after publication. You cannot prove cause.
• Invented percentages. Never quote figures about "how much more likely" a page type is cited unless you ran the test and show the sample. Vague industry numbers without a method are a reputational risk.
• "Special AI markup will do it." Google's guide to generative AI features says you do not need llms.txt files or other special markup to appear in them, because Google Search does not use them. Check the current guide before repeating any such claim.
• Mass-produced pages written for every query variation. That is the scaled content pattern Google warns against.

If a client has been sold AI search services elsewhere, the tutorial on getting a business recommended by AI search gives them a sober view of what is realistic.

Reporting results to the client

Report in plain words, with the method visible. A paragraph that works:

In our test (20 prompts, three runs each, one assistant, run 12 March), the
group was named in 12 of 60 answers, up from 5 of 60 in January. Three of
the seven new appearances cite the local news piece published on 20 Feb.
We cannot show the piece caused this, and results vary between runs. Next
test: 23 April. Next action: pitch two specialist titles with the group's
own planting data.

It has a sample size, a date, a comparison with last time, a limit and a next step. It makes no promise. If a client wants a framework behind the numbers, AMEC, the PR measurement body, has published GEO Principles and a practitioner's guide to GEO measurement. They assess AI-led discovery across three areas: upstream reputation (coverage, reviews, expert content), search and content readiness, and downstream AI outputs, meaning how the organisation appears in answers, including accuracy, prominence, framing, citations and omissions. They also warn against vanity metrics and false precision. Mapping your test to those three headings gives a client report a recognised structure: coverage won sits upstream, the crawler and Search Console checks are readiness, and the citation counts and error log are outputs. Pair the report with the regular coverage report, with a short section on citations, rather than creating a separate document the client has to learn to read.

Five prompts, two citations and an overconfident claim

Picture an agency (illustrative) that runs a one-off test with five prompts, sees the client cited twice, and tells the client "we're doing well in ChatGPT". A month later the client asks a colleague to try the same prompts and sees nothing. The sample was tiny, the runs were not repeated, and the colleague's search, location and history differ. The fix is to redo the work with 20 prompts and three runs, report a range, and change the wording to "named in 12 of 60 answers". The second report takes longer to write and lasts far longer. If you would rather design the test with someone, my AI implementation consultation covers that on a 1:1 call.

Questions PR teams ask about AI search citations

Can a PR agency guarantee a brand gets cited?

No. Answers vary between runs, users and locations, and no outside party controls what a model cites. You can improve the odds by earning coverage on sources the tools already use for that topic, and you can measure the result, but a guarantee would be a claim you cannot support.

Is blocking AI crawlers bad for PR?

It can be if the client wants to appear in ChatGPT search answers. OpenAI says a site must allow its search crawler, OAI-SearchBot, to be eligible. Its training crawler, GPTBot, is a separate robots.txt setting, so a client can allow search while refusing training. Ask the web team which is blocked and why; OpenAI says changes take about 24 hours to register.

How many prompts should a citation test use?

Start with 15 to 25 prompts that real customers might type, each run three times, because answers differ between runs. That gives 45 to 75 answers, enough to see patterns in which sources appear, without taking more than a few hours.

Do I need a paid tracking tool?

Not to start. A spreadsheet, a fixed prompt list and a person pasting results works for one or two clients. Paid tools help when you track many clients or prompts every week, so decide after your first manual test shows you what you want to measure.

Further reads

• Semrush AI Visibility vs Ahrefs Brand Radar: Which Suits an Agency? — Tool options for tracking AI visibility at scale.
• Google AI SEO Agency: Questions to Ask Before You Hire One — Questions to put to anyone selling AI search services.
• How to Choose an SEO Agency That Understands AI Search — How clients judge agencies on AI search work.
• AI PR Pitch Emails: Personalise for Journalists With a Review Step — Pitch wording for the placements that earn citations.
• AI Media List Tools: Muck Rack vs Semrush AI PR Toolkit vs ChatGPT — Building target lists of the publications you choose.
• AI SEO for Small Business: A Monthly Routine You Can Run Yourself — A monthly routine for the client's own site.
• AI Visibility for Hotels: Get Recommended in ChatGPT Answers — Free checks and a fix list for a small hotel: what ChatGPT and Google AI say about you today, where they get it, and what you can and cannot change.
• How Does ChatGPT Find a Business? What a Professional Firm Can Control — ChatGPT recommends firms from web search and public pages. Here is what it reads, what you control, and a simple test to see where you stand.
• How to Improve AI SEO: Getting Your Firm Cited in AI Answers — A small-firm checklist for being cited in ChatGPT and Google AI answers, with checks you can do yourself and a way to test with prompts.
• How to Monitor Brand Mentions in ChatGPT for a Client, Week by Week — A fixed prompt set, repeat runs, a log sheet and a weekday schedule for tracking a client's ChatGPT mentions, with tool costs and wrong-answer fixes.
• AI Agents for PR: What They Do and How an Agency Supervises Them — A PR agency supervision method for media lists, monitoring and pitch drafts, with a strict approval-before-send gate.
• Search Console and AI Overviews: What Agencies Can Report to Clients — What Google's Generative AI performance report and Bing's AI Performance report can honestly tell a client, plus a monthly reporting template.
• Using AI in Advertising: What a Small Agency Can Hand to a Tool — A task-by-task map of what a small ad agency can hand to AI, where each approval gate sits, and the jobs that stay human.
• Best AI Brand Visibility Software: Is AI Mentioning Your Business? — Check by hand first. Paid AI visibility trackers start around $29 a month, and only some small businesses need one. Prices, limits and a test.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Help Center (Searching the web with ChatGPT) and OpenAI crawler documentation; Google Search Central guides to succeeding in generative AI features and to AI features and your website; Search Console Help (Search generative AI control, Generative AI performance report); AMEC GEO Principles page; all checked 4 November 2026. All test figures are illustrative.